# François Chollet

François Chollet is a French AI researcher who created the Keras deep learning library and the ARC-AGI benchmark, and is a prominent critic of the claim that scaling large language models alone leads to general intelligence.

Francois Chollet is a French software engineer and AI researcher best known for creating [Keras](https://www.wikiprompt.org/wiki/keras), one of the most widely used [deep-learning](https://www.wikiprompt.org/wiki/deep-learning) libraries, and for designing the ARC-AGI benchmark, a widely cited test intended to measure general reasoning ability rather than memorized skill.

Chollet worked for many years at Google, where he developed Keras starting in 2015 as a high-level, user-friendly interface for building neural networks, initially running on top of [TensorFlow](https://www.wikiprompt.org/wiki/tensorflow) and later adopted as TensorFlow's official high-level API. Keras became known for making deep learning far more approachable to practitioners without deep systems expertise, contributing significantly to the field's rapid growth in the mid-2010s alongside [pytorch](https://www.wikiprompt.org/wiki/pytorch).

## ARC-AGI and the critique of scaling

In 2019 Chollet published the paper "On the Measure of Intelligence," which argued that most AI benchmarks conflate skill at a narrow task with intelligence, and proposed a definition of intelligence centered on skill-acquisition efficiency: how well a system can learn new tasks from very little data. The paper introduced the [arc-agi](https://www.wikiprompt.org/wiki/arc-agi) benchmark, a set of visual puzzle-like reasoning tasks that are simple for most humans but had proven very difficult for [machine-learning](https://www.wikiprompt.org/wiki/machine-learning) systems, including large [language models](https://www.wikiprompt.org/wiki/large-language-model), to solve without extensive task-specific training. Chollet has argued that ARC-AGI resists the memorization and pattern-matching shortcuts that let models perform well on many other benchmarks, and has been a prominent public skeptic of claims that scaling [pretraining](https://www.wikiprompt.org/wiki/pretraining) compute and data alone will produce [artificial-general-intelligence](https://www.wikiprompt.org/wiki/artificial-general-intelligence).

In 2024, Chollet co-launched the ARC Prize, a public competition with a cash prize for the first system to reach human-level performance on the benchmark, alongside entrepreneur Mike Knoop. The competition drew entries using both large-scale [reasoning models](https://www.wikiprompt.org/wiki/reasoning-model) and program-synthesis approaches, and provided some of the most closely watched public evidence in the debate over how close current systems are to general intelligence.

## Later work

Chollet left Google in 2024 and co-founded Ndea, a research lab pursuing program synthesis and other approaches to AI that Chollet has argued are more promising for genuine generalization than further scaling of [transformer](https://www.wikiprompt.org/wiki/transformer)-based models alone. He remains an active public commentator, frequently engaging in debates over [emergent-abilities](https://www.wikiprompt.org/wiki/emergent-abilities), [scaling-laws](https://www.wikiprompt.org/wiki/scaling-laws), and the meaning of benchmark results in AI research.

---
Source: https://www.wikiprompt.org/wiki/francois-chollet
License: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/)
Last updated: 2026-09-02T20:32:21.832591+00:00